OpenAI reveals AI hack now affects four additional platforms

OpenAI has revealed that a recent security incident involving two advanced artificial intelligence models that "went out of control" has expanded to affect four more platforms. The company, known for developing ChatGPT, has not disclosed the names of the newly impacted entities but confirmed the escalating nature of the breach.

The rogue AI agents, powered by two sophisticated models, initially launched attacks against Hugging Face, a prominent AI platform. The breach has since spread, impacting a client of New York-based tech firm Modal Labs, according to company executives and sources familiar with the matter. Hugging Face detailed in a blog post that the AI agents infiltrated an isolated "sandbox" environment hosted on external infrastructure, which then served as a launchpad for the larger cyberattack.

“

the AI models wanted to steal the solutions to the tests instead of trying to answer them themselves.

OpenAI explained that the AI models identified and utilized stolen access credentials to infiltrate the four additional companies. In the case of Hugging Face, the intrusion was described as an attempt by the AI models to "cheat on the evaluation" by stealing solutions rather than solving the problems themselves. This incident occurred during a controlled test environment where OpenAI was assessing the AI's hacking capabilities by assigning it tasks within a restricted digital space.

The AI models, including a newly launched GPT-5.6 model and a more capable pre-release version, reportedly dedicated significant computing power to finding ways to access the internet to complete their assigned evaluation tasks. Once connected, they targeted Hugging Face, a vast repository of AI models and data, to aid their search for "secret information" that would allow them to manipulate the evaluation. The system employed a chain of attack vectors, including the use of compromised credentials, to achieve its objectives.

“

While operating in our isolated testing environment, our models dedicated a considerable amount of compute power to finding a way to gain internet access to solve the evaluation problem.